BriefTechNews
Nvidia's Board Hunts for Huang's Replacement as Agents Take Over AWS

Nvidia's Board Hunts for Huang's Replacement as Agents Take Over AWS

7 min read · 14 sources

TL;DR
  • Nvidia's board is working with an outside firm to find a CEO replacement for Jensen Huang, with an interim chief under consideration.
  • AWS Emulator gives coding agents a local AWS copy to test against, reducing reliance on live cloud resources.
  • OpenAI's proof of the quantum Yang-Mills theory was formally verified in Lean by a committee led by physicist David Gross.
  • Google's Gemini Decisions API scores up to 1,000 questions per pass with 150-250 ms latency at $0.10 per 2M input tokens.
  • Direct agent work beat subagent delegation in a test: 67 seconds and $0.38 versus 123 seconds and $0.57 with Opus 5.5.

Nvidia’s board has hired an outside firm to vet CEO candidates as it decides whether Jensen Huang stays. Directors met Wednesday to weigh his fate after reports of extramarital affairs, one with a woman suspected of spying for China. The board is considering naming an interim chief while it searches for a permanent one, and the Wall Street Journal reports it’s preparing to oust him. Nvidia hasn’t commented.

This is not a normal succession story. Huang isn’t retiring; he’s the guy who turned Nvidia into a $3 trillion company. Leadership instability here doesn’t just shake one company - it shakes the entire AI hardware roadmap. Every engineer who’s planned a GPU cluster around Nvidia’s release cadence should watch this closely. A new CEO could mean a different strategic direction, different product priorities, and a very different conversation about who you’re buying silicon from in 2027.

Agentic browser traffic grew 8x between August and October, with Muse driving 60% of that activity.

AWS Emulator Gives Agents a Sandbox That Isn't the Cloud

Source: allthingsdistributed.com ↗

AWS has launched AWS Emulator, a tool that gives coding agents a local copy of AWS to test against. The post, authored by AWS CTO Werner Vogels and guest-written by Swami Sivasubramanian, traces its roots to early distributed systems work, including the Dynamo database and the Alf lock service built on Paxos.

The pitch is simple: agents write more code, so they need a reliable test environment that doesn’t burn money on live cloud resources. The emulator lets agents build and check most software locally, leaving engineers to decide what gets built and when the work is actually correct. The code is public.

For engineers, this changes the cost equation of agent-driven development. Instead of spinning up real AWS resources for every test run, agents can iterate against a local copy. That’s faster iteration, lower costs, and fewer surprises when something breaks in production because the agent never touched a real environment.

OpenAI's Shepherds Verify the Yang-Mills Proof

OpenAI’s AI-generated proof of the quantum Yang-Mills theory’s existence has been formally verified in the Lean programming language. A committee led by physicist David Gross released its findings on November 10, addressing one of the Clay Mathematics Institute’s Millennium Prize Problems.

The proof is too complex for humans to easily grasp - the committee doesn’t claim to fully understand it. That’s the point. This is a milestone in AI’s ability to tackle advanced mathematical problems, but the “shepherding” review process has drawn criticism from mathematicians who prefer open, community-based verification. OpenAI created the committee after backlash over its October release of 700+ AI-generated proofs in a single day.

For engineers, this matters because it’s a concrete demonstration that AI can produce verifiable results in domains where human review is the bottleneck. The Lean verification is a pattern worth stealing: formal verification as the acceptance test for AI-generated work.

The Gemini Decisions API Scores 1,000 Questions in One Pass

Source: blog.google ↗

Google has introduced the Gemini Decisions API, powered by the new Gemini Decision 1 model. It’s designed for fast, closed-set decisions like tool selection and policy checks, processing multimodal inputs (text, images, audio, video, PDFs) up to 256K tokens.

The headline numbers: it scores up to 1,000 questions in a single forward pass, returns probabilities with 150-250 ms end-to-end latency, and costs $0.10 per 2M input tokens with zero output tokens billed. That’s a fundamentally different cost model from token-by-token generation.

For engineers building decision systems, this is a cheaper, faster alternative for bounded tasks. You get built-in confidence scores for routing ambiguous cases, which means you can build a triage layer that sends only uncertain decisions to a full LLM. The tradeoff: it’s closed-set, so it won’t write you an essay - but it will tell you which of 1,000 policies applies to a given input.

Subagents Are Slower and Pricier Than Doing It Directly

Source: nikvdp.com ↗

Nik Vdp’s test with Opus 5.5 shows the cost of delegation. Using a subagent via summary took 123 seconds and cost $0.57. The main agent doing the task directly took 67 seconds and $0.38. The subagent approach was nearly twice as slow and 50% more expensive.

The problem is lossy summaries. When the main agent delegates, it compresses context, and the subagent works from that compressed version, leading to incomplete or incorrect work. The author suggests forked subagents, which copy the full conversation, as a better alternative - though forking gets expensive as context grows.

For engineers designing agent workflows, this is a concrete cautionary tale. Delegation isn’t free parallelism; it’s a context-compression tradeoff. The solution presented in the post is a fork that keeps the parts of a conversation it needs and drops the rest, which is a middle ground worth testing.

Who Captures the Margin in AI?

Source: tomtunguz.com ↗

Tom Tunguz’s analysis of margin capture across the AI value chain uses dollars per watt per year. At a 40% application gross margin, the pool splits as 3% power, 5% data center, 28% neocloud, 24% inference, and 40% application.

The interesting part is where pricing power lives. Compute prices are rising - Anthropic’s burst compute is at $13/GPU-hour, and CoreWeave’s new contracts are at $20-25/w/yr. CoreWeave acquiring Weights & Biases shows the vertical integration play. For engineers, this shapes cost structures across the stack: if you’re building on neoclouds, expect prices to climb as they capture more margin.

The B2A Economy Is Here

Source: a16z.news ↗

The a16z piece declares the rise of the business-to-agent economy, where AI agents are the primary internet users. Evidence: Shopify’s AI-driven traffic grew 8x year-over-year, and agentic browser traffic grew 8x between August and October, with Muse driving 60% of activity.

For engineers, this means building products that treat agents as first-class customers. That means agent-friendly APIs, machine-readable checkout flows, and content that doesn’t require human interaction to parse. Ignoring agent traffic risks obsolescence as more of the web’s activity shifts to autonomous actors.

The Answer to AI Is Engineering

Source: ghuntley.com ↗

Geoff Huntley’s argument is that the answer to AI security challenges is engineering, not more AI. Recent incidents - OpenCode’s breach, a PHP vulnerability, a GitLab RCE flaw, a certificate authority key that was just “4” - show the stakes. Agents writing software trained on vulnerable code will accelerate security issues.

The counterintuitive point: optimizing for incident response times also speeds up builds. The same infrastructure that lets you respond to a breach in minutes lets your CI pipeline ship faster. For engineers, this is a practical argument for investing in observability and response tooling - it pays off in both security and developer velocity.

The AI Talent War Gets Absurd

Source: wsj.com ↗

The WSJ reports that AI startup Serval is offering a $30,000 signing bonus and a $10,000 vacation to compete for software engineers and deployment strategists. This is the market telling you where the leverage is: people who can deploy and operate AI systems, not just prompt them.

The Levels of Agentic Engineering

Source: zachwills.net ↗

Zach Wills’s framework for leading AI transformations in R&D teams is built on one question: “What are you still doing yourself?” Each level involves removing yourself from manual loop parts, and the next level is found by shipping, not mandating.

The walls are infrastructure (CI, deploy pipelines) and policies (review requirements), which slow agent-driven work. The practical approach - shipping greenfield prototypes and real roadmap items - reveals how to integrate agents into existing workflows without a big-bang rewrite.

The Fall of Skills, the Rise of Networks

Source: jasonwei.net ↗

Jason Wei’s argument is that success in AI is shifting from “what you can do” to “who you know,” estimating that skills now account for 55% of success versus 75% five years ago. Three reasons: commoditized intelligence via coding agents, resource-rich labs having outsized advantages (hundreds of GPUs for RL runs), and reduced observability as companies keep models closed-source.

For engineers, this means networking and resource access are becoming as critical as technical ability for career advancement and research impact. The best strategy, per the post, is to be your best and truest self - which is both good advice and a reminder that the field is now social as much as technical.

Get the brief

Liked this one? Pick the beats you care about and get just those, five minutes a morning. Free, no hype, one click to leave.

About Me Author

My name is

BriefTechNews

A daily digest of what actually moved in AI, tech, crypto and fintech, assembled and written with AI, with every item linked to its source. Read More
Tags

You May Also Like